Designation: E178 − 08
Standard Practice for
Dealing With Outlying Observations1
This standard is issued under the fixed designation E178; the number immediately following the designation indicates the year of
original adoption or, in the case of revision, the year of last revision. A number in parentheses indicates the year of last reapproval. A
superscript epsilon (´) indicates an editorial change since the last revision or reapproval.
1. Scope
1.1 This practice covers outlying observations in samples
and how to test the statistical significance of them. An outlying
observation, or “outlier,” is one that appears to deviate mark-
edly from other members of the sample in which it occurs. In
this connection, the following two alternatives are of interest:
1.1.1 An outlying observation may be merely an extreme
manifestation of the random variability inherent in the data. If
this is true, the value should be retained and processed in the
same manner as the other observations in the sample.
1.1.2 On the other hand, an outlying observation may be the
result of gross deviation from prescribed experimental proce-
dure or an error in calculating or recording the numerical value.
In such cases, it may be desirable to institute an investigation
to ascertain the reason for the aberrant value. The observation
may even actually be rejected as a result of the investigation,
though not necessarily so. At any rate, in subsequent data
analysis the outlier or outliers will be recognized as probably
e
being from a different population than that of the other sample
values.
gP
1.2 It is our purpose here to provide statistical rules that will
lead the experimenter almost unerringly to look for causes of
outliers when they really exist, and hence to decide whether
alternative 1.1.1 above, is not the more plausible hypothesis to
accept, as compared to alternative 1.1.2, in order that the most
appropriate action in further data analysis may be taken. The
En
procedures covered herein apply primarily to the simplest kind
of experimental data, that is, replicate measurements of some
property of a given material, or observations in a supposedly
single random sample. Nevertheless, the tests suggested do
cover a wide enough range of cases in practice to have broad
utility.
1
This practice is under the jurisdiction of ASTM Committee E11 on Quality and
Statistics and is the direct responsibility of Subcommittee E11.10 on Sampling /
Statistics.
Current edition approved Oct. 1, 2008. Published November 2008. Originally
approved in 1961. Last previous edition approved in 2002 as E178 – 02. DOI:
10.1520/E0178-08.
Copyright © ASTM International, 100 Barr Harbor Drive, PO Box C700, West Conshohocken, PA 19428-2959. United States
An American National Standard
2. Referenced Documents
2.1 ASTM Standards:2
E456 Terminology Relating to Quality and Statistics
3. Terminology
3.1 Definitions: The terminology defined in Terminology
E456 applies to this standard unless modified herein.
ir
3.1.1 outlier—see outlying observation.
3.1.2 outlying observation, n—an observation that appears
to deviate markedly in value from other members of the sample
a.
in which it appears.
4. Significance and Use
di 4.1 When the experimenter is clearly aware that a gross
deviation from prescribed experimental procedure has taken
place, the resultant observation should be discarded, whether or
not it agrees with the rest of the data and without recourse to
statistical tests for outliers. If a reliable correction procedure,
for example, for temperature, is available, the observation may
sometimes be corrected and retained.
4.2 In many cases evidence for deviation from prescribed
procedure will consist primarily of the discordant value itself.
In such cases it is advisable to adopt a cautious attitude. Use of
one of the criteria discussed below will sometimes permit a
clear-cut decision to be made. In doubtful cases the experi-
menter’s judgment will have considerable influence. When the
experimenter cannot identify abnormal conditions, he should at
least report the discordant values and indicate to what extent
they have been used in the analysis of the data.
4.3 Thus, for purposes of orientation relative to the over-all
problem of experimentation, our position on the matter of
screening samples for outlying observations is precisely the
following:
4.3.1 Physical Reason Known or Discovered for Outlier(s):
2
For referenced ASTM standards, visit the ASTM website, www.astm.org, or
contact ASTM Customer Service at service@astm.org. For Annual Book of ASTM
Standards volume information, refer to the standard’s Document Summary page on
the ASTM website.
1
 E178 − 08
4.3.1.1 Reject observation(s).
4.3.1.2 Correct observation(s) on physical grounds.
4.3.1.3 Reject it (them) and possibly take additional obser-
vation(s).
4.3.2 Physical Reason Unknown—Use Statistical Test:
4.3.2.1 Reject observation(s).
4.3.2.2 Correct observation(s) statistically.
4.3.2.3 Reject it (them) and possibly take additional obser-
vation(s).
4.3.2.4 Employ truncated-sample theory for censored obser-
vations.
4.4 The statistical test may always be used to support a
judgment that a physical reason does actually exist for an
outlier, or the statistical criterion may be used routinely as a
basis to initiate action to find a physical cause.
5. Basis of Statistical Criteria for Outliers
5.1 There are a number of criteria for testing outliers. In all
of these, the doubtful observation is included in the calculation
of the numerical value of a sample criterion (or statistic), which
is then compared with a critical value based on the theory of
random sampling to determine whether the doubtful observa-
a.
tion is to be retained or rejected. The critical value is that value
of the sample criterion which would be exceeded by chance
with some specified (small) probability on the assumption that
all the observations did indeed constitute a random sample
from a common system of causes, a single parent population,
distribution or universe. The specified small probability is
di
called the “significance level” or “percentage point” and can be
e
thought of as the risk of erroneously rejecting a good obser-
vation. It becomes clear, therefore, that if there exists a real
shift or change in the value of an observation that arises from
gP
nonrandom causes (human error, loss of calibration of
instrument, change of measuring instrument, or even change of
time of measurements, etc.), then the observed value of the
sample criterion used would exceed the “critical value” based
on random-sampling theory. Tables of critical values are
usually given for several different significance levels, for
En
example, 5 %, 1 %. For statistical tests of outlying
observations, it is generally recommended that a low signifi-
cance level, such as 1 %, be used and that significance levels
greater than 5 % should not be common practice.
NOTE 1—In this practice, we will usually illustrate the use of the 5 %
significance level. Proper choice of level in probability depends on the
particular problem and just what may be involved, along with the risk that
one is willing to take in rejecting a good observation, that is, if the
null-hypothesis stating “all observations in the sample come from the
same normal population” may be assumed correct.
5.2 It should be pointed out that almost all criteria for
outliers are based on an assumed underlying normal (Gaussian)
population or distribution. When the data are not normally or
approximately normally distributed, the probabilities associ-
ated with these tests will be different. Until such time as criteria
not sensitive to the normality assumption are developed, the
experimenter is cautioned against interpreting the probabilities
too literally.
5.3 Although our primary interest here is that of detecting
outlying observations, we remark that some of the statistical
2
criteria presented may also be used to test the hypothesis of
normality or that the random sample taken did come from a
normal or Gaussian population. The end result is for all
practical purposes the same, that is, we really wish to know
whether we ought to proceed as if we have in hand a sample of
homogeneous normal observations.
6. Recommended Criteria for Single Samples
6.1 Let the sample of n observations be denoted in order of
increasing magnitude by x1 ≤ x2 ≤ x3 ≤ ... ≤ x n. Let xn be the
doubtful value, that is the largest value. The test criterion, Tn,
recommended here for a single outlier is as follows:
T n 5 ~ x n 2 x̄ ! /s (1)
where:
x̄ = arithmetic average of all n values, and
s = estimate of the population standard deviation based on
the sample data, calculated as follows:
ir
s = n n
! ( ~ x i 2x̄ ! 2
! (x 2 2
i 2n·x̄
i51
5
i51
n21 n21
n
S( D
n 2
5!( i51
x i 22
n21
i51
xi /n
If x1 rather than xn is the doubtful value, the criterion is as
follows:
T 1 5 ~ x̄ 2 x 1 ! /s (2)
The critical values for either case, for the 1 and 5 % levels of
significance, are given in Table 1. Table 1 and the following
tables give the “one-sided” significance levels. In the previous
tentative recommended practice (1961), the tables listed values
of significance levels double those in the present practice, since
it was considered that the experimenter would test either the
lowest or the highest observation (or both) for statistical
significance. However, to be consistent with actual practice and
in an attempt to avoid further misunderstanding, single-sided
significance levels are tabulated here so that both viewpoints
can be represented.
6.2 The hypothesis that we are testing in every case is that
all observations in the sample come from the same normal
population. Let us adopt, for example, a significance level of
0.05. If we are interested only in outliers that occur on the high
side, we should always use the statistic Tn = (xn − x̄)/s and take
as critical value the 0.05 point of Table 1. On the other hand,
if we are interested only in outliers occurring on the low side,
we would always use the statistic T1 = (x̄ − x1)/s and again take
as a critical value the 0.05 point of Table 1. Suppose, however,
that we are interested in outliers occurring on either side, but
do not believe that outliers can occur on both sides simultane-
ously. We might, for example, believe that at some time during
the experiment something possibly happened to cause an
extraneous variation on the high side or on the low side, but
that it was very unlikely that two or more such events could
have occurred, one being an extraneous variation on the high
side and the other an extraneous variation on the low side. With
 E178 − 08
TABLE 1 Critical Values for T (One-Sided Test) When Standard Deviation is Calculated from the Same SampleA
Number of Upper 0.1 % Upper 0.5 % Upper 1 % Upper 2.5 % Upper 5 % Upper 10 %
Observations, Significance Significance Significance Significance Significance Significance
n Level Level Level Level Level Level
3 1.155 1.155 1.155 1.155 1.153 1.148
4 1.499 1.496 1.492 1.481 1.463 1.425
5 1.780 1.764 1.749 1.715 1.672 1.602

6 2.011 1.973 1.944 1.887 1.822 1.729

7 2.201 2.139 2.097 2.020 1.938 1.828
8 2.358 2.274 2.221 2.126 2.032 1.909
9 2.492 2.387 2.323 2.215 2.110 1.977
10 2.606 2.482 2.410 2.290 2.176 2.036

11 2.705 2.564 2.485 2.355 2.234 2.088

12 2.791 2.636 2.550 2.412 2.285 2.134
13 2.867 2.699 2.607 2.462 2.331 2.175
14 2.935 2.755 2.659 2.507 2.371 2.213
15 2.997 2.806 2.705 2.549 2.409 2.247

16 3.052 2.852 2.747 2.585 2.443 2.279

17 3.103 2.894 2.785 2.620 2.475 2.309
18 3.149 2.932 2.821 2.651 2.504 2.335
19 3.191 2.968 2.854 2.681 2.532 2.361
20 3.230 3.001 2.884 2.709 2.557 2.385

ir
21 3.266 3.031 2.912 2.733 2.580 2.408
22 3.300 3.060 2.939 2.758 2.603 2.429
23 3.332 3.087 2.963 2.781 2.624 2.448

a.
24 3.362 3.112 2.987 2.802 2.644 2.467
25 3.389 3.135 3.009 2.822 2.663 2.486

26 3.415 3.157 3.029 2.841 2.681 2.502

27 3.440 3.178 3.049
di 2.859 2.698 2.519
28 3.464 3.199 3.068 2.876 2.714 2.534
29 3.486 3.218 3.085 2.893 2.730 2.549
30 3.507 3.236 3.103 2.908 2.745 2.563

31 3.528 3.253 3.119 2.924 2.759 2.577

32 3.546 3.270 3.135 2.938 2.773 2.591
e
33 3.565 3.286 3.150 2.952 2.786 2.604
34 3.582 3.301 3.164 2.965 2.799 2.616
35 3.599 3.316 3.178 2.979 2.811 2.628
gP

36 3.616 3.330 3.191 2.991 2.823 2.639

37 3.631 3.343 3.204 3.003 2.835 2.650
38 3.646 3.356 3.216 3.014 2.846 2.661
39 3.660 3.369 3.228 3.025 2.857 2.671
40 3.673 3.381 3.240 3.036 2.866 2.682

41 3.687 3.393 3.251 3.046 2.877 2.692

En

42 3.700 3.404 3.261 3.057 2.887 2.700

43 3.712 3.415 3.271 3.067 2.896 2.710
44 3.724 3.425 3.282 3.075 2.905 2.719
45 3.736 3.435 3.292 3.085 2.914 2.727

46 3.747 3.445 3.302 3.094 2.923 2.736

47 3.757 3.455 3.310 3.103 2.931 2.744
48 3.768 3.464 3.319 3.111 2.940 2.753
49 3.779 3.474 3.329 3.120 2.948 2.760
50 3.789 3.483 3.336 3.128 2.956 2.768

51 3.798 3.491 3.345 3.136 2.964 2.775

52 3.808 3.500 3.353 3.143 2.971 2.783
53 3.816 3.507 3.361 3.151 2.978 2.790
54 3.825 3.516 3.368 3.158 2.986 2.798
55 3.834 3.524 3.376 3.166 2.992 2.804

56 3.842 3.531 3.383 3.172 3.000 2.811

57 3.851 3.539 3.391 3.180 3.006 2.818
58 3.858 3.546 3.397 3.186 3.013 2.824
59 3.867 3.553 3.405 3.193 3.019 2.831
60 3.874 3.560 3.411 3.199 3.025 2.837

61 3.882 3.566 3.418 3.205 3.032 2.842

62 3.889 3.573 3.424 3.212 3.037 2.849
63 3.896 3.579 3.430 3.218 3.044 2.854

3
 E178 − 08
TABLE 1 Continued
Number of Upper 0.1 % Upper 0.5 % Upper 1 % Upper 2.5 % Upper 5 % Upper 10 %
Observations, Significance Significance Significance Significance Significance Significance
n Level Level Level Level Level Level
64 3.903 3.586 3.437 3.224 3.049 2.860
65 3.910 3.592 3.442 3.230 3.055 2.866

66 3.917 3.598 3.449 3.235 3.061 2.871

67 3.923 3.605 3.454 3.241 3.066 2.877
68 3.930 3.610 3.460 3.246 3.071 2.883
69 3.936 3.617 3.466 3.252 3.076 2.888
70 3.942 3.622 3.471 3.257 3.082 2.893

71 3.948 3.627 3.476 3.262 3.087 2.897

72 3.954 3.633 3.482 3.267 3.092 2.903
73 3.960 3.638 3.487 3.272 3.098 2.908
74 3.965 3.643 3.492 3.278 3.102 2.912
75 3.971 3.648 3.496 3.282 3.107 2.917

76 3.977 3.654 3.502 3.287 3.111 2.922

77 3.982 3.658 3.507 3.291 3.117 2.927
78 3.987 3.663 3.511 3.297 3.121 2.931
79 3.992 3.669 3.516 3.301 3.125 2.935
80 3.998 3.673 3.521 3.305 3.130 2.940

ir
81 4.002 3.677 3.525 3.309 3.134 2.945
82 4.007 3.682 3.529 3.315 3.139 2.949
83 4.012 3.687 3.534 3.319 3.143 2.953
84 4.017 3.691 3.539 3.323 3.147 2.957

a.
85 4.021 3.695 3.543 3.327 3.151 2.961

86 4.026 3.699 3.547 3.331 3.155 2.966

87 4.031 3.704 3.551 3.335 3.160 2.970
88 4.035 3.708 3.555
di 3.339 3.163 2.973
89 4.039 3.712 3.559 3.343 3.167 2.977
90 4.044 3.716 3.563 3.347 3.171 2.981

91 4.049 3.720 3.567 3.350 3.174 2.984

92 4.053 3.725 3.570 3.355 3.179 2.989
e
93 4.057 3.728 3.575 3.358 3.182 2.993
94 4.060 3.732 3.579 3.362 3.186 2.996
95 4.064 3.736 3.582 3.365 3.189 3.000
gP

96 4.069 3.739 3.586 3.369 3.193 3.003

97 4.073 3.744 3.589 3.372 3.196 3.006
98 4.076 3.747 3.593 3.377 3.201 3.011
99 4.080 3.750 3.597 3.380 3.204 3.014
100 4.084 3.754 3.600 3.383 3.207 3.017

101 4.088 3.757 3.603 3.386 3.210 3.021

102 4.092 3.760 3.607 3.390 3.214 3.024
En

103 4.095 3.765 3.610 3.393 3.217 3.027

104 4.098 3.768 3.614 3.397 3.220 3.030
105 4.102 3.771 3.617 3.400 3.224 3.033

106 4.105 3.774 3.620 3.403 3.227 3.037

107 4.109 3.777 3.623 3.406 3.230 3.040
108 4.112 3.780 3.626 3.409 3.233 3.043
109 4.116 3.784 3.629 3.412 3.236 3.046
110 4.119 3.787 3.632 3.415 3.239 3.049

111 4.122 3.790 3.636 3.418 3.242 3.052

112 4.125 3.793 3.639 3.422 3.245 3.055
113 4.129 3.796 3.642 3.424 3.248 3.058
114 4.132 3.799 3.645 3.427 3.251 3.061
115 4.135 3.802 3.647 3.430 3.254 3.064

116 4.138 3.805 3.650 3.433 3.257 3.067

117 4.141 3.808 3.653 3.435 3.259 3.070
118 4.144 3.811 3.656 3.438 3.262 3.073
119 4.146 3.814 3.659 3.441 3.265 3.075
120 4.150 3.817 3.662 3.444 3.267 3.078

121 4.153 3.819 3.665 3.447 3.270 3.081

122 4.156 3.822 3.667 3.450 3.274 3.083
123 4.159 3.824 3.670 3.452 3.276 3.086
124 4.161 3.827 3.672 3.455 3.279 3.089

4
 E178 − 08
TABLE 1 Continued
Number of Upper 0.1 % Upper 0.5 % Upper 1 % Upper 2.5 % Upper 5 % Upper 10 %
Observations, Significance Significance Significance Significance Significance Significance
n Level Level Level Level Level Level
125 4.164 3.831 3.675 3.457 3.281 3.092
126 4.166 3.833 3.677 3.460 3.284 3.095
127 4.169 3.836 3.680 3.462 3.286 3.097
128 4.173 3.838 3.683 3.465 3.289 3.100
129 4.175 3.840 3.686 3.467 3.291 3.102
130 4.178 3.843 3.688 3.470 3.294 3.104

131 4.180 3.845 3.690 3.473 3.296 3.107

132 4.183 3.848 3.693 3.475 3.298 3.109
133 4.185 3.850 3.695 3.478 3.302 3.112
134 4.188 3.853 3.697 3.480 3.304 3.114
135 4.190 3.856 3.700 3.482 3.306 3.116

136 4.193 3.858 3.702 3.484 3.309 3.119

137 4.196 3.860 3.704 3.487 3.311 3.122
138 4.198 3.863 3.707 3.489 3.313 3.124
139 4.200 3.865 3.710 3.491 3.315 3.126
140 4.203 3.867 3.712 3.493 3.318 3.129

141 4.205 3.869 3.714 3.497 3.320 3.131

142 4.207 3.871 3.716 3.499 3.322 3.133

ir
143 4.209 3.874 3.719 3.501 3.324 3.135
144 4.212 3.876 3.721 3.503 3.326 3.138
145 4.214 3.879 3.723 3.505 3.328 3.140

a.
146 4.216 3.881 3.725 3.507 3.331 3.142
147 4.219 3.883 3.727 3.509 3.334 3.144
Tn = (xn − x̄)/s

ΠΠΠSo D
n n n n 2

o
i51
s x i 2x̄ d 2 ox
i51
i
2
2n·x̄ 2
ox
i51
i
2
2
i51
xi /n
5 5

A
n21 n21 n21
T1 = [(x̄ − x1)/s]x1≤ x2 ≤ ... ≤ xn
Values of T are taken from Ref (1). All values have been adjusted for division by n – 1 instead of n in calculating s.
e di
this point of view we should use the statistic T n = (xn − x̄)/s or against the doubtful value having come from the same popu-
the statistic T1 = (x̄ − x1)/ s whichever is larger. If in this lation as the others (assuming the population is normally
gP

instance we use the 0.05 point of Table 1 as our critical value, distributed). Investigation of the doubtful value is therefore
the true significance level would be twice 0.05 or 0.10. If we indicated.
wish a significance level of 0.05 and not 0.10, we must in this 6.3 An alternative system, the Dixon criteria, based entirely
case use as a critical value the 0.025 point of Table 1. Similar on ratios of differences between the observations is described
considerations apply to the other tests given below. in the literature (2)3 and may be used in cases where it is
6.2.1 Example 1—As an illustration of the use of Tn and desirable to avoid calculation of s or where quick judgment is
En

Table 1, consider the following ten observations on breaking called for. For the Dixon test, the sample criterion or statistic
strength (in pounds) of 0.104-in. hard-drawn copper wire: 568, changes with sample size. Table 2 gives the appropriate
570, 570, 570, 572, 572, 572, 578, 584, 596. See Fig. 1. The statistic to calculate and also gives the critical values of the
doubtful observation is the high value, x10 = 596. Is the value statistic for the 1, 5, and 10 % levels of significance.
of 596 significantly high? The mean is x̄ = 575.2 and the 6.3.1 Example 2—As an illustration of the use of Dixon’s
estimated standard deviation is s = 8.70. We compute test, consider again the observations on breaking strength given
T 10 5 ~ 596 2 575.2! /8.70 5 2.39 (3) in Example 1, and suppose that a large number of such samples
had to be screened quickly for outliers and it was judged too
From Table 1, for n = 10, note that a T10 as large as 2.39
time-consuming to compute s. Table 2 indicates use of
would occur by chance with probability less than 0.05. In fact,
so large a value would occur by chance not much more often r 11 5 ~ x n 2 x n21 ! / ~ x n 2 x 2! (4)
than 1 % of the time. Thus, the weight of the evidence is
Thus, for n = 10,
r 11 5 ~ x 10 2 x 9 ! / ~ x 10 2 x 2 ! (5)

For the measurements of breaking strength above,

r 11 5 ~ 596 2 584! / ~ 596 2 570! 5 0.462 (6)

FIG. 1 Ten Observations of Breaking Strength from Example 1 in 3

The boldface numbers in parentheses refer to the list of references at the end of
6.2.1 this practice.

5
 E178 − 08
TABLE 2 Dixon Criteria for Testing of Extreme Observation (Single Sample)A
Significance Level (One-Sided Test)
n Criterion
10 percent 5 percent 1 percent
3 r10 = (x2 − x1)/(xn − x1) if smallest value is suspected; 0.886 0.941 0.988
4 = (xn − xn−1)/(xn − x1) if largest value is suspected 0.679 0.765 0.889
5 0.557 0.642 0.780
6 0.482 0.560 0.698
7 0.434 0.507 0.637
8 r11 = (x2 − x1)/(xn−1 − x1) if smallest value is suspected; 0.479 0.554 0.683
9 = (xn − xn−1)/(xn − x2) if largest value is suspected. 0.441 0.512 0.635
10 0.409 0.477 0.597
11 r21 = (x3 − x1)/(xn−1 − x1) if smallest value is suspected; 0.517 0.576 0.679
12 = (xn − xn−2)/(xn − x2) if largest value is suspected. 0.490 0.546 0.642
13 0.467 0.521 0.615
14 r22 = (x3 − x1)/(xn−2 − x1) if smallest value is suspected; 0.492 0.546 0.641
15 = (xn − xn−2)/(xn − x3) if largest value is suspected. 0.472 0.525 0.616
16 0.454 0.507 0.595
17 0.438 0.490 0.577
18 0.424 0.475 0.561
19 0.412 0.462 0.547
20 0.401 0.450 0.535
21 0.391 0.440 0.524
22 0.382 0.430 0.514
23 0.374 0.421 0.505
24 0.367 0.413 0.497

ir
25 0.360 0.406 0.489
26 0.354 0.399 0.486
27 0.348 0.393 0.475
28 0.342 0.387 0.469

a.
29 0.337 0.381 0.463
30 0.332 0.376 0.457
A
x1 # x2 # ... # xn. (See Ref (2), Appendix.)

which is a little less than 0.477, the 5 % critical value for

n = 10. Under the Dixon criterion, we should therefore not
di residuals test of Tietjen and Moore (7) could be used. An
example in astronomy follows.
consider this observation as an outlier at the 5 % level of 6.5.1 Example 3—There is one rather famous set of obser-
e
significance. These results illustrate how borderline cases may vations that a number of writers on the subject of outlying
be accepted under one test but rejected under another. It should observations have referred to in applying their various tests for
be remembered, however, that the T-statistic discussed above is “outliers.” This classic set consists of a sample of 15 observa-
gP

the best one to use for the single-outlier case, and final
statistical judgment should be based on it. See Ferguson (3,4).
6.3.2 Further examination of the sample observations on
breaking strength of hand-drawn copper wire indicates that
none of the other values need testing.
En
tions of the vertical semidiameters of Venus made by Lieuten-
ant Herndon in 1846 (8). In the reduction of the observations,
Prof. Pierce assumed two unknown quantities and found the
following residuals which have been arranged in ascending
order of magnitude:

NOTE 2—With experience we may usually just look at the sample −1.40 in. −0.24 −0.05 0.18 0.48
values to observe if an outlier is present. However, strictly speaking the −0.44 −0.22 0.06 0.20 0.63
−0.30 −0.13 0.10 0.39 1.01
statistical test should be applied to all samples to guarantee the signifi-
cance levels used. Concerning “multiple” tests on a single sample, we See Fig. 2.
comment on this below. The deviations − 1.40 and 1.01 appear to be outliers. Here
6.4 A test equivalent to Tn (or T1) based on the sample sum the suspected observations lie at each end of the sample. Much
of squared deviations from the mean for all the observations less work has been accomplished for the case of outliers at both
and the sum of squared deviations omitting the “outlier” is ends of the sample than for the case of one or more outliers at
given by Grubbs (5). only one end of the sample. This is not necessarily because the
6.5 The next type of problem to consider is the case where “one-sided” case occurs more frequently in practice but be-
we have the possibility of two outlying observations, the least cause “two-sided’’ tests are much more difficult to deal with.
and the greatest observation in a sample. (The problem of For a high and a low outlier in a single sample, we give two
testing the two highest or the two lowest observations is procedures below, the first being a combination of tests, and the
considered below.) In testing the least and the greatest obser- second a single test of Tietjen and Moore (7) which may have
vations simultaneously as probable outliers in a sample, we use nearly optimum properties. For optimum procedures when
the ratio of sample range to sample standard deviation test of there is an independent estimate at hand, s2 or σ 2, see (9).
David, Hartley, and Pearson (6). The significance levels for this 6.6 For the observations on the semi-diameter of Venus
sample criterion are given in Table 3. Alternatively, the largest given above, all the information on the measurement error is

6
 E178 − 08
TABLE 3 Critical Values (One-Sided Test) for w/s (Ratio of Range or more outliers. The lowest measurement, − 1.40 in., is 1.418
to Sample Standard Deviation)A below the sample mean, and the highest measurement, 1.01 in.,
5 Percent 1 Percent 0.5 Percent is 0.992 above the mean. Since these extremes are not
Number of
Significance Significance Significance
Observations, n symmetric about the mean, either both extremes are outliers, or
Level Level Level
3 2.00 2.00 2.00 else only − 1.40 is an outlier. That − 1.40 is an outlier can be
4 2.43 2.44 2.45 verified by use of the T1 statistic. We have
5 2.75 2.80 2.81
6 3.01 3.10 3.12 T 1 5 ~ x̄ 2 x 1 ! /s 5 @ 0.018 2 ~ 21.40! # /0.551 5 2.574 (10)
7 3.22 3.34 3.37
8 3.40 3.54 3.58 This value is greater than the critical value for the 5 % level,
9 3.55 3.72 3.77 2.409 from Table 1, so we reject − 1.40. Since we have decided
10 3.68 3.88 3.94
11 3.80 4.01 4.08 that − 1.40 should be rejected, we use the remaining 14
12 3.91 4.13 4.21 observations and test the upper extreme 1.01, either with the
13 4.00 4.24 4.32 criterion
14 4.09 4.34 4.43
15 4.17 4.43 4.53 T n 5 ~ x n 2 x̄ ! /s (11)
16 4.24 4.51 4.62
17 4.31 4.59 4.69 or with Dixon’s r22. Omitting − 1.40 and renumbering the
18 4.38 4.66 4.77
19 4.43 4.73 4.84
observations, we compute
20 4.49 4.79 4.91 x̄ 5 1.67/14 5 0.119, s 5 0.401, (12)
30 4.89 5.25 5.39
40 5.15 5.54 5.69
and

ir
50 5.35 5.77 5.91
60 5.50 5.93 6.09 T 14 5 ~ 1.01 2 0.119! /0.401 5 2.22 (13)
80 5.73 6.18 6.35
100 5.90 6.36 6.54 From Table 1, for n = 14, we find that a value as large as 2.22

a.
150 6.18 6.64 6.84
200 6.38 6.85 7.03 would occur by chance more than 5 % of the time, so we
500 6.94 7.42 7.60 should retain the value 1.01 in further calculations. We next
1000 7.33 7.80 7.99 calculate
A
See Ref (6), where:
w 5 x n 5 x1

x1 # x2 # { # xn
di r 22 5 ~ x 14 2 x 12! / ~ x 14 2 x 3 !
5 ~ 1.01 2 0.48! / ~ 1.0110.24!
50.53/1.25
(14)

ΠΠΠSo D 50.424
n n n n 2

o ox ox
e
s x i 2x̄ d 2 i
2
2n·x̄ 2
i
2
2 xi /n
i51 i51 i51 i51
s5 5 5 From Table 2 for n = 14, we see that the 5 % critical value
n21 n21 n21
for r22 is 0.546. Since our calculated value (0.424) is less than
gP

the critical value, we also retain 1.01 by Dixon’s test, and no

contained in the sample of 15 residuals. In cases like this, further values would be tested in this sample.
where no independent estimate of variance is available (that is,
NOTE 3—It should be noted that in repeated application of outlier tests
we still have the single sample case), a useful statistic is the to a sample, the overall significance level changes. If we apply k tests, an
ratio of the range of the observations to the sample standard acceptable rule would be to use a significance level of α/k for each test so
deviation: that the overall significance level will be approximately α.
En

w/s 5 ~ x n 2 x 1 ! /s (7) 6.8 For suspected observations on both the high and low
sides in the sample, and to deal with the situation in which
where:
some of k ≥ 2 suspected outliers are larger and some smaller
s5 =( @ ~ x 2 x̄ ! / ~ n 2 1 ! #
i
2
(8) than the remaining values in the sample, Tietjen and Moore (7)
suggest the following statistic. Let the sample values be x1, x2,
If xn is about as far above the mean, x̄, as x1 is below x̄, and x3, ... xn and compute the sample mean, x̄. Then compute the n
if w/s exceeds some chosen critical value, then one would absolute residuals
conclude that both the doubtful values are outliers. If, however,
x1 and xn are displaced from the mean by different amounts, ? ? ? ?
r 1 5 x 1 , 2, x̄ , r 2 5 x 2 , 2, x̄ , … r n 5 x n , 2, x̄ ? ? (15)
some further test would have to be made to decide whether to Now relabel the original observations x1, x2, ... , xn as z’s in
reject as outlying only the lowest value or only the highest such a manner that zi is that x whose ri is the ith smallest
value or both the lowest and highest values. absolute residual above. This now means that z1 is that
6.7 For this example the mean of the deviations is x̄ = 0.018, observation x which is closest to the mean and that zn is the
s = 0.551, and observation x which is farthest from the mean. The Tietjen-
Moore statistic for testing the significance of the k largest
w/s 5 @ 1.01 2 ~ 21.40! # /0.551 5 2.41/0.551 5 4.374 (9)
residuals is then
From Table 3 for n = 15, we see that the value of w/s
= 4.374 falls between the critical values for the 1 and 5 %
levels, so if the test were being run at the 5 % level of
Ek 5 F( ~n2k

i51
z i 2 z̄ k ! 2 /
n

( ~ z 2 z̄ !
i51
i
2
G (16)

significance, we would conclude that this sample contains one where:

7
 E178 − 08

FIG. 2 Fifteen Residuals from the Semidiameters of Venus from

Example 3 in 6.5.1

n2k
z̄ k 5 ( z /~n 2 k!
i51
i (17)
is the mean of the (n − k) least extreme observations and z is
the mean of the full sample.
6.8.1 Applying this test to the above data, we find that the
total sum of squares of deviations for the entire sample is
4.24964. Omitting —1.40 and 1.01, the suspected two outliers,
we find that the sum of squares of deviations for the reduced
sample of 13 observations is 1.24089. Then
E2 = 1.24089 ⁄4.24964 = 0.292, and by using Table 4, we find
that this observed E2 is slightly smaller than the 5 % critical
based on the ratio of the sample sum of squares when the two
doubtful values are omitted to the sample sum of squares when
the two doubtful values are included. If simplicity in calcula-
tion is the prime requirement, then the Dixon type of test
(actually omitting one observation in the sample) might be
used for this case. In illustrating the test procedure, we give the
following Examples 4 and 5.
6.9.1 Example 4—In a comparison of strength of various
plastic materials, one characteristic studied was the percentage
elongation at break. Before comparison of the average elonga-
tion of the several materials, it was desirable to isolate for

ir
value of 0.317, so that the E2 test would reject both of the further study any pieces of a given material which gave very
observations, —1.40 and 1.01. We would probably take this small elongation at breakage compared with the rest of the
latter recommendation, since the level of significance for the pieces in the sample. In this example, one might have primary

E2 test is precisely 0.05 whereas that for the double application
of a test for a single outlier cannot be guaranteed to be smaller
than 1 — (0.95)2 = 0.0975. The table of percentage points of Ek
was computed by Monte Carlo methods on a high-speed
electronic calculator.
6.9 We next turn to the case where we may have the two
largest or the two smallest observations as probable outliers.
e
Here, we employ a test provided by Grubbs (5, 10) which is
a.
interest only in outliers to the left of the mean for study, since
very high readings indicate exceeding plasticity, a desirable
characteristic.
di 6.9.1.1 Ten measurements of percentage elongation at break
made on material No. 23 follow: 3.73, 3.59, 3.94, 4.13, 3.04,
2.22, 3.23, 4.05, 4.11, and 2.02. See Fig. 3. Arranged in
ascending order of magnitude, these measurements are: 2.02,
2.22, 3.04, 3.23, 3.59, 3.73, 3.94, 4.05, 4.11, 4.13. The

TABLE 4 1000 X Tietjen-Moore Critical Values (One-Sided Test) for Ek

gP

k n
α 50 45 40 35 30 25 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3
1A 0.01 .748 .728 .704 .669 .624 .571 .499 .484 .459 .440 .422 .404 .374 .337 .311 .274 .235 .197 .156 .110 68 .29 .4 ...
0.05 .796 .776 .756 .732 .698 .654 .594 .579 .562 .544 .525 .503 .479 .453 .423 .390 .353 .310 .262 .207 .145 .81 .25 .1
0.10 .820 .802 .784 .762 .730 .692 .638 .624 .610 .593 .576 .556 .534 .510 .482 .451 .415 .374 .326 .270 .203 .127 .49 .3
2 0.01 .636 .607 .574 .533 .482 .418 .339 .323 .306 .290 .263 .238 .207 .181 .159 .134 .101 .78 .50 .28 .12 .2 ... ...
0.05 .684 .658 .629 .596 .549 .493 .416 .398 .382 .362 .340 .317 .293 .262 .234 .204 .172 .137 .99 .65 .34 .10 .1 ...
En

0.10 .708 .684 .657 .624 .582 .528 .460 .442 .424 .406 .384 .360 .337 .309 .278 .250 .214 .175 .137 .94 .56 .22 .2 ...
3 0.01 .550 .518 .480 .435 .386 .320 .236 .219 .206 .188 .166 .146 .123 .103 .83 .64 .44 .26 .14 .6 .1 ... ... ...
.0.05 .599 .567 .534 .495 .443 .381 .302 .287 .267 .248 .227 .206 .179 .156 .133 .107 .83 .57 .34 .16 .4 ... ... ...
.0.10 .622 .593 .562 .523 .475 .417 .338 .322 .304 .284 .263 .240 .216 .189 .162 .138 .108 .80 .53 .27 .9 ... ... ...
4 0.01 .482 .446 .408 .364 .308 .245 .170 .156 .141 .122 .107 .90 .72 .56 .42 .30 .18 .9 .4 ... ... ... ... ...
0.05 .529 .492 .458 .417 .364 .298 .221 .203 .187 .170 .153 .134 .112 .92 .73 .55 .37 .21 .10 ... ... ... ... ...
0.10 .552 .522 .486 .443 .391 .331 .252 .234 .217 .198 .182 .160 .138 .116 .94 .73 .52 .32 .16 ... ... ... ... ...
5 0.01 .424 .386 .347 .299 .250 .188 .121 .108 .94 .79 .68 .54 .42 .31 .20 .12 .6 ... ... ... ... ... ... ...
0.05 .468 .433 .395 .351 .298 .236 .163 .146 .132 .116 .102 .84 .68 .53 .39 .26 .14 ... ... ... ... ... ... ...
0.10 .492 .459 .422 .379 .325 .264 .188 .172 .156 .140 .122 .105 .86 .68 .52 .36 .22 ... ... ... ... ... ... ...
6 0.01 .376 .336 .298 .252 .204 .146 .86 .74 .62 .52 .40 .32 .22 .14 .8 ... ... ... ... ... ... ... ... ...
0.05 .417 .381 .343 .298 .246 .186 .119 .105 .91 .78 .67 .52 .39 .28 .18 ... ... ... ... ... ... ... ... ...
0.10 .440 .406 .367 .324 .270 .210 .138 .124 .110 .95 .82 .67 .52 .38 .26 ... ... ... ... ... ... ... ... ...
7 0.01 .334 .294 .258 .211 .166 .110 .58 .50 .41 .32 .24 .18 .12 ... ... ... ... ... ... ... ... ... ... ...
0.05 .373 .337 .297 .254 .203 .146 .85 .74 .62 .50 .41 .30 .21 ... ... ... ... ... ... ... ... ... ... ...
0.10 .396 .360 .320 .276 .224 .168 .102 .89 .76 .64 .53 .40 .29 ... ... ... ... ... ... ... ... ... ... ...
8 0.01 .297 .258 .220 .177 .132 .87 .40 .32 .26 .18 .14 ... ... ... ... ... ... ... ... ... ... ... ... ...
0.05 .334 .299 .259 .214 .166 .114 .59 .50 .41 .32 .24 ... ... ... ... ... ... ... ... ... ... ... ... ...
0.10 .355 .320 .278 .236 .186 .132 .72 .62 .51 .42 .32 ... ... ... ... ... ... ... ... ... ... ... ... ...
9 0.01 .264 .228 .190 .149 .108 .66 .26 .20 .14 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
0.05 .299 .263 .223 .181 .137 .89 .41 .33 .26 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
0.10 .319 .284 .243 .202 .154 .103 .51 .42 .34 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
10 0.01 .235 .200 .164 .124 .87 .50 .17 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
0.05 .268 .233 .195 .154 .112 .68 .28 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
0.10 .287 .252 .212 .172 .126 .80 .35 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
A
From Grubbs (1950, Table 1) for n # 25.

8
 E178 − 08
are the two smallest ranges, 4420 and 4549. For testing these
two suspected outliers, the statistic S1,22/S2 of Table 5 is
probably the best to use.
NOTE 4—Kudo (11) indicates that if the two outliers are due to a shift
in location or level, as compared to the scale σ, then the optimum sample
criterion for testing should be of the type:
min (2x̄ − xi − xj)/ s = (2x̄ − x1 − x2)/s in our Example 5.
6.9.2.2 The distances arranged in increasing order of mag-
FIG. 3 Ten Measurements of Percentage Elongation at Break nitude are:
from Example 4 in 6.9.1 4420 4782
4549 4803
4730 4833
questionable readings are the two lowest, 2.02 and 2.22. We 4765 4838
can test these two low readings simultaneously by using the The value of S2 is 158 592. Omission of the two shortest
following criterion of Table 5: ranges, 4420 and 4549, and recalculation, gives S1,22 equal to
S 1,2 2 /S 2
(18) 8590.8. Thus,
S 1,2 2 /S 2 5 8590.8/158 592 5 0.054 (23)
For the above measurements:
which is significant at the 0.01 level (See Table 5). It is thus
S 2 5( ~ x 2 x̄ !
i51
i
2
highly unlikely that the two shortest ranges (occurring actually

ir
2
5@n( x 2 ~( x ! i
2
# /n i (19) from excessive yaw) could have come from the same popula-
5 @ 10~ 121.3594! 2 ~ 34.06! 2 # /10 tion as that represented by the other six ranges. It should be
55.351, noted that the critical values in Table 5 for the 1 % level of

a.
significance are smaller than those for the 5 % level. So for this
and particular test, the calculated value is significant if it is less
n
than the chosen critical value.
S 1,2 2 5 ( ~ x 2 x̄
i53
i 1,2 !2 di 6.10 By Monte Carlo methods using an electronic
F
5 ~n 2 2! (x
i53
n

i
2
2 S( D G ~i53
n

x i
2

/ n 2 2!
(20)
calculator, Tietjen and Moore (7) extended the tables of
percentage points for the two highest or the two lowest
2
5 @ 8 ~ 112.3506! 2 ~ 29.82! # /8 observations to k > 2 highest or lowest sample values. Their
e
59.5724/8 results are given in Table 6 where
51.197 n2k n n2k

Lk 5 ( ~ x 2 x̄ ! / ( ~ x
2
2 x̄ ! 2
and x̄ k 5 ( x /~n 2 k!.
F ( x /~n 2 2!G
n i k i i
gP

i51 i51 i51

where x̄ 1,2 5 i (21)
i53 Note that their L2 equals our Sn,n−12/S2. For k = 1, their
We find: critical values agreed with exact values calculated by Grubbs
(1950). This new table may be used to advantage in many
S 1,2 2 /S 2 5 1.197/5.351 5 0.224 (22)
practical problems of interest.
From Table 5 for n = 10, the 5 % significance level for S1,2 6.11 If simplicity in calculation is very important, or if a
En

2
2/S is 0.2305. Since the calculated value is less than the critical
value, we should conclude that both 2.02 and 2.22 are outliers.
In a situation such as the one described in this example, where
the outliers are to be isolated for further analysis, a significance
level as high as 5 % or perhaps even 10 % would probably be
used in order to get a reasonable size of sample for additional
study. The problem may really be one of economics, and we
use probability as a sensible basis for action.
6.9.2 Example 5—The following ranges (horizontal dis-
tances in yards from gun muzzle to point of impact of a
projectile) were obtained in firings from a weapon at a constant
angle of elevation and at the same weight of charge of
propellant powder.
Distances in Yards
4782
4838
4765
4549
6.9.2.1 It is desired to make a judgment on whether the
projectiles exhibit uniformity in ballistic behavior or if some of
the ranges are inconsistent with the others. The doubtful values
large number of samples must be examined individually for
outliers, the questionable observations may be tested with the
application of Dixon’s criteria. Disregarding the lowest range,
4420, we test if the next lowest range, 4549, is outlying. With
n = 7, we see from Table 2 that r10 is the appropriate statistic.
Renumbering the ranges as xi to x7, beginning with 4549, we
find:
r 10 5 ~ x 2 2 x 1 ! / ~ x 7 2 x 1 !
5 ~ 4730 2 4549! / ~ 4838 2 4549!
(24)
5181/289
50.626
which is only a little less than the 1 % critical value, 0.637,
for n = 7. So, if the test is being conducted at any significance
4420
4803 level greater than a 1 % level, we would conclude that 4549 is
4730 an outlier. Since the lowest of the original set of ranges, 4420,
4833 is even more outlying than the one we have just tested, it can
be classified as an outlier without further testing. We note here,
however, that this test did not use all of the sample observa-
tions.

9
 E178 − 08
2
TABLE 5 Critical Values for S n−1,n/ S2, orS21,2/S2 for Simultaneously Testing the Two Largest or Two Smallest ObservationsA
Number of Lower 0.1 % Lower 0.5 % Lower 1 % Lower 2.5 % Lower 5 % Lower 10 %
Observations, Significance Significance Significance Significance Significance Significance
n Level Level Level Level Level Level
4 0.0000 0.0000 0.0000 0.0002 0.0008 0.0031
5 0.0003 0.0018 0.0035 0.0090 0.0183 0.0376

6 0.0039 0.0116 0.0186 0.0349 0.0564 0.0920

7 0.0135 0.0308 0.0440 0.0708 0.1020 0.1479
8 0.0290 0.0563 0.0750 0.1101 0.1478 0.1994
9 0.0489 0.0851 0.1082 0.1492 0.1909 0.2454
10 0.0714 0.1150 0.1414 0.1864 0.2305 0.2863

11 0.0953 0.1448 0.1736 0.2213 0.2667 0.3227

12 0.1198 0.1738 0.2043 0.2537 0.2996 0.3552
13 0.1441 0.2016 0.2333 0.2836 0.3295 0.3843
14 0.1680 0.2280 0.2605 0.3112 0.3568 0.4106
15 0.1912 0.2530 0.2859 0.3367 0.3818 0.4345

16 0.2136 0.2767 0.3098 0.3603 0.4048 0.4562

17 0.2350 0.2990 0.3321 0.3822 0.4259 0.4761
18 0.2556 0.3200 0.3530 0.4025 0.4455 0.4944
19 0.2752 0.3398 0.3725 0.4214 0.4636 0.5113
20 0.2939 0.3585 0.3909 0.4391 0.4804 0.5270

21 0.3118 0.3761 0.4082 0.4556 0.4961 0.5415

ir
22 0.3288 0.3927 0.4245 0.4711 0.5107 0.5550
23 0.3450 0.4085 0.4398 0.4857 0.5244 0.5677
24 0.3605 0.4234 0.4543 0.4994 0.5373 0.5795

a.
25 0.3752 0.4376 0.4680 0.5123 0.5495 0.5906

26 0.3893 0.4510 0.4810 0.5245 0.5609 0.6011

27 0.4027 0.4638 0.4933 0.5360 0.5717 0.6110
28 0.4156 0.4759 0.5050 0.5470 0.5819 0.6203
29 0.4279 0.4875 0.5162 0.5574 0.5916 0.6292
30

31
0.4397

0.4510
0.4985

0.5091
di
0.5268

0.5369
0.5672

0.5766
0.6008

0.6095
0.6375

0.6455
32 0.4618 0.5192 0.5465 0.5856 0.6178 0.6530
33 0.4722 0.5288 0.5557 0.5941 0.6257 0.6602
e
34 0.4821 0.5381 0.5646 0.6023 0.6333 0.6671
35 0.4917 0.5469 0.5730 0.6101 0.6405 0.6737
gP

36 0.5009 0.5554 0.5811 0.6175 0.6474 0.6800

37 0.5098 0.5636 0.5889 0.6247 0.6541 0.6860
38 0.5184 0.5714 0.5963 0.6316 0.6604 0.6917
39 0.5266 0.5789 0.6035 0.6382 0.6665 0.6972
40 0.5345 0.5862 0.6104 0.6445 0.6724 0.7025

41 0.5422 0.5932 0.6170 0.6506 0.6780 0.7076

42 0.5496 0.5999 0.6234 0.6565 0.6834 0.7125
En

43 0.5568 0.6064 0.6296 0.6621 0.6886 0.7172

44 0.5637 0.6127 0.6355 0.6676 0.6936 0.7218
45 0.5704 0.6188 0.6412 0.6728 0.6985 0.7261

46 0.5768 0.6246 0.6468 0.6779 0.7032 0.7304

47 0.5831 0.6303 0.6521 0.6828 0.7077 0.7345
48 0.5892 0.6358 0.6573 0.6876 0.7120 0.7384
49 0.5951 0.6411 0.6623 0.6921 0.7163 0.7422
50 0.6008 0.6462 0.6672 0.6966 0.7203 0.7459

51 0.6063 0.6512 0.6719 0.7009 0.7243 0.7495

52 0.6117 0.6560 0.6765 0.7051 0.7281 0.7529
53 0.6169 0.6607 0.6809 0.7091 0.7319 0.7563
54 0.6220 0.6653 0.6852 0.7130 0.7355 0.7595
55 0.6269 0.6697 0.6894 0.7168 0.7390 0.7627

56 0.6317 0.6740 0.6934 0.7205 0.7424 0.7658

57 0.6364 0.6782 0.6974 0.7241 0.7456 0.7687
58 0.6410 0.6823 0.7012 0.7276 0.7489 0.7716
59 0.6454 0.6862 0.7049 0.7310 0.7520 0.7744
60 0.6497 0.6901 0.7086 0.7343 0.7550 0.7772

61 0.6539 0.6938 0.7121 0.7375 0.7580 0.7798

62 0.6580 0.6975 0.7155 0.7406 0.7608 0.7824
63 0.6620 0.7010 0.7189 0.7437 0.7636 0.7850
64 0.6658 0.7045 0.7221 0.7467 0.7664 0.7874

10
 E178 − 08
TABLE 5 Continued
Number of Lower 0.1 % Lower 0.5 % Lower 1 % Lower 2.5 % Lower 5 % Lower 10 %
Observations, Significance Significance Significance Significance Significance Significance
n Level Level Level Level Level Level
65 0.6696 0.7079 0.7253 0.7496 0.7690 0.7898
66 0.6733 0.7112 0.7284 0.7524 0.7716 0.7921
67 0.6770 0.7144 0.7314 0.7551 0.7741 0.7944
68 0.6805 0.7175 0.7344 0.7578 0.7766 0.7966
69 0.6839 0.7206 0.7373 0.7604 0.7790 0.7988
70 0.6873 0.7236 0.7401 0.7630 0.7813 0.8009

71 0.6906 0.7265 0.7429 0.7655 0.7836 0.8030

72 0.6938 0.7294 0.7455 0.7679 0.7859 0.8050
73 0.6970 0.7322 0.7482 0.7703 0.7881 0.8070
74 0.7000 0.7349 0.7507 0.7727 0.7902 0.8089
75 0.7031 0.7376 0.7532 0.7749 0.7923 0.8108

76 0.7060 0.7402 0.7557 0.7772 0.7944 0.8127

77 0.7089 0.7427 0.7581 0.7794 0.7964 0.8145
78 0.7117 0.7453 0.7605 0.7815 0.7983 0.8162
79 0.7145 0.7477 0.7628 0.7836 0.8002 0.8180
80 0.7172 0.7501 0.7650 0.7856 0.8021 0.8197

81 0.7199 0.7525 0.7672 0.7876 0.8040 0.8213

82 0.7225 0.7548 0.7694 0.7896 0.8058 0.8230

ir
83 0.7250 0.7570 0.7715 0.7915 0.8075 0.8245
84 0.7275 0.7592 0.7736 0.7934 0.8093 0.8261
85 0.7300 0.7614 0.7756 0.7953 0.8109 0.8276

a.
86 0.7324 0.7635 0.7776 0.7971 0.8126 0.8291
87 0.7348 0.7656 0.7796 0.7989 0.8142 0.8306
88 0.7371 0.7677 0.7815 0.8006 0.8158 0.8321
89 0.7394 0.7697 0.7834 0.8023 0.8174 0.8335
90 0.7416 0.7717 0.7853
di 0.8040 0.8190 0.8349

91 0.7438 0.7736 0.7871 0.8057 0.8205 0.8362

92 0.7459 0.7755 0.7889 0.8073 0.8220 0.8376
93 0.7481 0.7774 0.7906 0.8089 0.8234 0.8389
94 0.7501 0.7792 0.7923 0.8104 0.8248 0.8402
e
95 0.7522 0.7810 0.7940 0.8120 0.8263 0.8414

96 0.7542 0.7828 0.7957 0.8135 0.8276 0.8427

97 0.7562 0.7845 0.7973 0.8149 0.8290 0.8439
gP

98 0.7581 0.7862 0.7989 0.8164 0.8303 0.8451

99 0.7600 0.7879 0.8005 0.8178 0.8316 0.8463
100 0.7619 0.7896 0.8020 0.8192 0.8329 0.8475

101 0.7637 0.7912 0.8036 0.8206 0.8342 0.8486

102 0.7655 0.7928 0.8051 0.8220 0.8354 0.8497
103 0.7673 0.7944 0.8065 0.8233 0.8367 0.8508
104 0.7691 0.7959 0.8080 0.8246 0.8379 0.8519
En

105 0.7708 0.7974 0.8094 0.8259 0.8391 0.8530

106 0.7725 0.7989 0.8108 0.8272 0.8402 0.8541

107 0.7742 0.8004 0.8122 0.8284 0.8414 0.8551
108 0.7758 0.8018 0.8136 0.8297 0.8425 0.8563
109 0.7774 0.8033 0.8149 0.8309 0.8436 0.8571
110 0.7790 0.8047 0.8162 0.8321 0.8447 0.8581

111 0.7806 0.8061 0.8175 0.8333 0.8458 0.8591

112 0.7821 0.8074 0.8188 0.8344 0.8469 0.8600
113 0.7837 0.8088 0.8200 0.8356 0.8479 0.8610
114 0.7852 0.8101 0.8213 0.8367 0.8489 0.8619
115 0.7866 0.8114 0.8225 0.8378 0.8500 0.8628

116 0.7881 0.8127 0.8237 0.8389 0.8510 0.8637

117 0.7895 0.8139 0.8249 0.8400 0.8519 0.8646
118 0.7909 0.8152 0.8261 0.8410 0.8529 0.8655
119 0.7923 0.8164 0.8272 0.8421 0.8539 0.8664
120 0.7937 0.8176 0.8284 0.8431 0.8548 0.8672

121 0.7951 0.8188 0.8295 0.8441 0.8557 0.8681

122 0.7964 0.8200 0.8306 0.8451 0.8567 0.8689
123 0.7977 0.8211 0.8317 0.8461 0.8576 0.8697
124 0.7990 0.8223 0.8327 0.8471 0.8585 0.8705
125 0.8003 0.8234 0.8338 0.8480 0.8593 0.8713

11
 E178 − 08
TABLE 5 Continued
Number of Lower 0.1 % Lower 0.5 % Lower 1 % Lower 2.5 % Lower 5 % Lower 10 %
Observations, Significance Significance Significance Significance Significance Significance
n Level Level Level Level Level Level
126 0.8016 0.8245 0.8348 0.8490 0.8602 0.8721
127 0.8028 0.8256 0.8359 0.8499 0.8611 0.8729
128 0.8041 0.8267 0.8369 0.8508 0.8619 0.8737
129 0.8053 0.8278 0.8379 0.8517 0.8627 0.8744
130 0.8065 0.8288 0.8389 0.8526 0.8636 0.8752

131 0.8077 0.8299 0.8398 0.8535 0.8644 0.8759

132 0.8088 0.8309 0.8408 0.8544 0.8652 0.8766
133 0.8100 0.8319 0.8418 0.8553 0.8660 0.8773
134 0.8111 0.8329 0.8427 0.8561 0.8668 0.8780
135 0.8122 0.8339 0.8436 0.8570 0.8675 0.8787

136 0.8134 0.8349 0.8445 0.8578 0.8683 0.8794

137 0.8145 0.8358 0.8454 0.8586 0.8690 0.8801
138 0.8155 0.8368 0.8463 0.8594 0.8698 0.8808
139 0.8166 0.8377 0.8472 0.8602 0.8705 0.8814
140 0.8176 0.8387 0.8481 0.8610 0.8712 0.8821

141 0.8187 0.8396 0.8489 0.8618 0.8720 0.8827

142 0.8197 0.8405 0.8498 0.8625 0.8727 0.8834
143 0.8207 0.8414 0.8506 0.8633 0.8734 0.8840

ir
144 0.8218 0.8423 0.8515 0.8641 0.8741 0.8846
145 0.8227 0.8431 0.8523 0.8648 0.8747 0.8853

146 0.8237 0.8440 0.8531 0.8655 0.8754 0.8859

a.
147 0.8247 0.8449 0.8539 0.8663 0.8761 0.8865
148 0.8256 0.8457 0.8547 0.8670 0.8767 0.8871
149 0.8266 0.8465 0.8555 0.8677 0.8774 0.8877
n
di S2 5 o s x 2 x̄ d
i51
i
2

x1 # x2 # ... # xn
n

S 2 1,2 5 o s x 2 x̄
i51
i 1,2 d2
e
n
1
x̄ 1,2 5
n22 ox
i53
i
gP

n22
2
S n21,n 5 oi51
s x i 2 x̄ n 2 1,n d 2

n22
1
x̄ n21,n 5
n22 o
i51
xi
En

A
These significance levels are taken from Table 11, Ref (1). An observed ratio less than the appropriate critical ratio in this table calls for rejection of the null hypothesis.

6.12 Rejection of Several Outliers— So far we have dis- the various rejection rules relative to changes in level or scale.
cussed procedures for detecting one or two outliers in the same For several outliers and repeated rejection of observations,
sample, but these techniques are not generally recommended Ferguson points out that the sample coefficient of skewness,
for repeated rejection, since if several outliers are present in the n
sample the detection of one or two spurious values may be =b 1 5 =n ( ~ x i 2 x̄ ! 3 / ~ n 2 1 ! 3/2 s 3
i51
(25)
“masked” by the presence of other anomalous observations.
Outlying observations occur due to a shift in level (or mean), n

or a change in scale (that is, change in variance of the 5 =n ( ~ x 2 x̄ ! / @ ( ~ x 2 x̄ !

i
3
i
2 3/2
#
i51

observations), or both. Ferguson (3,4) has studied the power of

12
 E178 − 08
TABLE 6 1000 X Tietjen-Moore Critical Values (One-Sided Test) for Lk
k n
α 50 45 40 35 30 25 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3
A
1 0.01 .768 .745 .722 .690 .650 .607 .539 .522 .504 .485 .463 .440 .414 .386 .355 .321 .283 .241 .195 .145 .93 .44 .10 ...
0.025 .796 .776 .756 .732 .699 .654 .594 .579 .562 .544 .525 .503 .479 .453 .423 .390 .353 .310 .262 .207 .145 .81 .25 .1
0.05 .820 .802 .784 .762 .730 .692 .638 .624 .610 .593 .576 .556 .534 .510 .482 .451 .415 .374 .326 .270 .203 .127 .49 3
0.10 .840 .826 .812 .792 .766 .732 .685 .673 .660 .646 .631 .613 .594 .573 .548 .520 .488 .450 .405 .350 .283 .199 .98 .11
2B 0.01 . 667 .641 .610 .573 .527 .468 .391 .373 .353 .332 .310 .286 .261 .233 .204 .174 .141 .108 .75 .44 .19 .4 ... ...
0.025 .697 .667 .644 .610 .567 .512 .439 .421 .403 .382 .360 .337 .311 .284 .254 .221 .186 .149 .110 .71 .35 .9 ... ...
0.05 .720 .698 .673 .641 .601 .550 .480 .464 .446 .426 .405 .382 .357 .330 .300 .267 .230 .191 .148 .102 .56 .18 .1 ...
0.10 .746 .726 .702 .674 .637 .591 .527 .511 .494 .476 .456 .435 .411 .384 .355 .323 .286 . 245 .199 .148 .92 .38 .3 ...
3 0.01 .592 .558 .522 .484 .434 .377 .300 .272 .260 .237 .219 .194 .172 .147 .120 .98 .70 .48 .28 .10 .2 ... ... ...
0.025 .622 .592 .561 .527 .479 .417 .341 .321 .299 . 282 .261 .239 .214 .184 .162 .129 .100 .73 .45 .21 .5 ... ... ...
0.05 .646 .618 .588 .554 .506 .450 .377 .354 .337 .322 .300 .276 .250 .224 .196 .162 .129 .99 .64 .32 .10 ... ... ...
0.10 .673 .648 .622 .586 .523 .489 .420 .398 .384 .364 .342 .322 .298 .270 .240 .208 .170 .134 .95 .56 .20 ... ... ...
4 0.01 .531 .498 .460 .418 .369 .308 .231 .211 .192 .171 .151 .132 .113 .94 .70 .52 .32 .18 .8 ... ... ... ... ...
0.025 .559 .529 .491 .455 .408 .342 .265 .243 .226 .208 .185 .167 .145 .122 .96 .74 .52 .30 .13 ... ... ... ... ...
0.05 .588 .556 .523 .482 .434 .374 .299 .277 .259 .240 .219 .197 .174 .150 .125 .98 .70 .45 .22 ... ... ... ... ...
0.10 .614 .586 .554 .516 .472 .412 .339 . 316 .302 .282 .260 .236 .212 .186 .159 .128 .98 .68 .38 ... ... ... ... ...
5 0.01 .483 .444 .408 .364 .312 .246 .175 .154 .140 .126 .108 .90 .72 .56 .38 .26 .12 ... ... ... ... ... ... ...
0.025 .510 .473 .433 .398 .352 .282 .209 .189 .171 .151 .135 .113 .95 .77 .57 .40 .23 ... ... ... ... ... ... ...
0.05 .535 .502 .468 .424 . 376 .312 .238 .217 .200 .181 .159 .140 .122 .98 .76 .54 .34 ... ... ... ... ... ... ...
0.10 .562 .533 .499 .458 .411 .350 .273 .251 .236 .216 .194 .172 .150 .126 .103 .74 .51 ... ... ... ... ... ... ...
6 0.01 .438 .399 .364 .321 .268 .204 .136 .118 .104 .91 .72 .57 .46 .33 .19 ... ... ... ... ... ... ... ... ...
0.025 .466 .430 .387 .348 .302 .233 .165 .145 .129 .117 .96 .78 .63 .47 .31 ... ... ... ... ... ... ... ... ...
0.05 .490 .456 .421 .376 .327 .262 .188 .168 .154 .136 .115 .97 .79 .60 .42 ... ... ... ... ... ... ... ... ...

ir
0.10 .518 .488 .451 .410 .359 .296 .220 .199 .184 .165 .144 .124 .104 .82 .62 ... ... ... ... ... ... ... ... ...
7 0.01 .400 .361 .324 .282 .229 .168 .104 .88 .76 .64 .49 .37 .27 ... ... ... ... ... ... ... ... ... ... ...
0.025 .428 .391 .348 .308 .261 .192 .128 .108 .95 .82 .65 .51 .38 ... ... ... ... ... ... ... ... ... ... ...

a.
0.05 .450 .417 .378 .334 . 283 .222 .150 .130 .116 .100 .82 .66 ,50 ... ... ... ... ... ... ... ... ... ... ...
0.10 .477 .447 .408 .365 .316 .251 .176 .158 .142 .125 .104 .86 .68 ... ... ... ... ... ... ... ... ... ...
8 0.01 .368 .328 .292 .250 .196 .144 .78 .64 .53 .44 .30 ... ... ... ... ... ... ... ... ... ... ... ... ...
0.025 .392 .356 .314 .274 . 226 .159 .98 .80 .68 .58 .45 ... ... ... ... ... ... ... ... ... ... ... ... ...
0.05 .414 .382 .342 .297 .245 .184 .115 .99 .86 .72 .55 ... ... ... ... ... ... ... ... ... ... ... ... ...
0.10 .442 .410 .372 .328 .276 .213 .140 .124 .108 .92 .73 ... ... ... ... ... ... ... ... ... ... ... ... ...
9 0.01
0.025
0.05
.336
.363
.383
.296
.325
.350
.262
.283
.310
.220
.242
.264
.166
.193
. 212
.112
.132
.154
.58
.73
.88
.46
.59
.74
.36
.48
.62
...
...
...
...
...
...
di
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
0.10 .410 .378 .338 .294 .240 .180 .110 .94 .80 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
10 0.01 .308 .270 .234 .194 . 142 .92 .42 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
e
0.025 .334 .295 .257 .213 .165 .108 .54 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
0.05 .356 .320 .280 .235 .183 .126 .66 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
0.10 .380 .348 .307 .262 .210 .152 .85 ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
gP

A
From Grubbs (1950, Table I) for n # 25.
B
From Grubbs (1972, Table II).

should be used for “one-sided” tests (change in level of 8, then the observation farthest from the mean is rejected and
several observations in the same direction), and the sample the same procedure repeated until no further sample values are
En

coefficient of kurtosis, judged as outliers. (As is well-known =b 1 and b2 are also used
n as tests of normality.)
b2 5 n ( ~x
i51
i 2 x̄ ! 4 / ~ n 2 1 ! 2 s 4 (26) NOTE 5—In the above equations for =b 1 and b2, s is defined as used
in this standard:

S( D
n n n n n 2

( ~ x 2 x̄ ! / @ ( ~ x 2 x̄ ! #2
! ! !
4 2
5n
i51
i i (
i51
~ x i 2x̄ ! 2 (x
i51
i
2
2n·x̄ 2
(x
i51
i
2
2
i51
xi /n
5 5
is recommended for “two-sided” tests (change in level to ~ n21 ! ~ n21 ! n21
higher and lower values) and also for changes in scale 6.12.1 The significance levels in Table 7 and Table 8 for
(variance) (see Note 5). In applying the above tests, the =b 1 or sample sizes of 5, 10, 15, and 20 (and 25 for b2) were obtained
the b2, or both, are computed and if their observed values by Ferguson on an IBM 704 computer using a sampling
exceed those for significance levels given in Table 7 and Table

TABLE 7 Significance Levels (One-Sided Test) for œ b 1 TABLE 8 Significance Levels (One-Sided Test) for b2
Signifi- Significance n
n
cance Level, percent 5A 10A 15A 20A 25A 50 75 100
Level, per- 1 3.11 4.83 5.08 5.23 5.00 4.88 4.59 4.39
5A 10 A
15 A
20 A
25 30 35 40 50 60
cent 5 2.89 3.85 4.07 4.15 4.00 3.99 3.87 3.77
1 1.34 1.31 1.20 1.11 1.06 0.98 0.92 0.87 0.79 0.72 A
These values were obtained by Ferguson, using a Monte Carlo procedure. For
5 1.05 0.92 0.84 0.79 0.71 0.66 0.62 0.59 0.53 0.49
n = 25; Ferguson’s Monte Carlo values of b2 agree with Pearson’s computed
A
These values were obtained by Ferguson, using a Monte Carlo procedure. values.

13
 E178 − 08
experiment or “Monte Carlo’’ procedure. The significance
levels for the other sample sizes are from Pearson, E. S. “Table
of Percentage Points of = b 1 and b2 in Normal Samples; a
Rounding Off,” Biometrika, Vol 52, 1965, pp. 282–285.
6.12.2 The = b 1 and b2 statistics have the optimum property
of being “locally” best against one-sided and two-sided
alternatives, respectively. The =b 1 test is good for up to 50 %
spurious observations in the sample for the one-sided case, and
the b2 test is optimum in the two-sided alternatives case for up
to 21 % “contamination” of sample values. For only one or two
outliers the sample statistics of the previous paragraphs are
recommended, and Ferguson (3) discusses in detail their
optimum properties of pointing out one or two outliers.
6.12.2.1 Instead of the more complicated =b 1 and b2
statistics, one can use Table 4 and Table 6 (7) for sample sizes
and percentage points given.
7. Recommended Criterion Using Independent Standard
ir
Deviation
7.1 Suppose that an independent estimate of the standard
deviation is available from previous data. This estimate may be
a.
from a single sample of previous similar data or may be the
result of combining estimates from several such previous sets
of data. In any event, each estimate is said to have degrees of
freedom equal to one less than the sample size that it is based
on. The proper combined estimate is a weighted average of the
several values of s2, the weights being proportional to the
di
respective degrees of freedom. The total degrees of freedom in
e
the combined estimate is then the sum of the individual degrees
of freedom. When one uses an independent estimate of the
standard deviation, sv, the test criterion recommended here for
gP
an outlier is as follows:
T' 1 5 ~ x̄ 2 x 1 ! /s (27)
or:
T' n 5 ~ x n 2 x̄ ! /s v (28)
En
where:
v = total number of degrees of freedom.
7.2 The critical values for T'1 and T'n for the 5 % and 1 %
significance levels are due to David (12) and are given in Table
9. In Table 9 the subscript v = df indicates the total number of
degrees of freedom associated with the independent estimate of
standard deviation σ and n indicates the number of observa-
tions in the sample under study. When Table 9 skips over
desired values of v and/or n, the user may interpolate from the
neighboring values. The user may also use the simpler and
more conservative approach and choose the larger of the
neighboring critical values. We illustrate with an example on
interlaboratory testing.
7.3 Example 6—Interlaboratory Testing—In an analysis of
interlaboratory test procedures, data representing normalities
of sodium hydroxide solutions were determined by twelve
different laboratories. In all the standardizations, a 0.1 N
sodium hydroxide solution was prepared by the Standard
Methods Committee using carbon-dioxide-free distilled water.
14
Potassium acid phthalate (P.A.P.), obtained from the National
Institute of Standards and Technology, was used as the test
standard.
7.3.1 Test data by the twelve laboratories are given in Table
10. The P.A.P. readings have been coded to simplify the
calculations. The variances between the three readings within
all laboratories were found to be homogeneous. A one-way
classification in the analysis of variance was first analyzed to
determine if the variation in laboratory results (averages) was
statistically significant. This variation was significant and
indicated a need for action, so tests for outliers were then
applied to isolate the particular laboratories whose results gave
rise to the significant variation.
7.3.2 Table 11 shows that the variation between laboratories
is highly significant. To test if this (very significant) variation
is due to one (or perhaps two) laboratories that obtained
“outlying” results (that is, perhaps showing nonstandard
technique), we can test the laboratory averages for outliers.
From the analysis of variance, we have an estimate of the
variance of an individual reading as 0.008793, based on 24
degrees of freedom. The estimated standard deviation of an
individual measurement is =0.00879350.094 and the estimated
standard deviation of the average of three readings is therefore
0.094/ =350.054.
7.3.3 Since the estimate of within-laboratory variation is
independent of any difference between laboratories, we can use
the statistic T'1 of 7.1 to test for outliers. An examination of the
deviations of the laboratory averages from the grand average
indicates that Laboratory 10 obtained an average reading much
lower than the grand average, and that Laboratory 12 obtained
a high average compared to the over-all average. To first test if
Laboratory 10 is an outlier, we compute
T' 5 ~ 1.871 2 0.745! /0.054 5 20.9 (29)
7.3.4 This value of T' is obviously significant at a very low
level of probability (P << 0.01—Refer to Table 9 with n = 12
and v = 24 degrees of freedom). We conclude, therefore, that
the test methods of Laboratory 10 should be investigated.
7.3.5 Excluding Laboratory 10, we compute a new grand
average of 1.973 and test if the results of Laboratory 12 are
outlying. We have
T' 5 ~ 2.327 2 1.973! /0.054 5 6.56 (30)
and this value of T' is significant at P<< 0.01 (T' = 20.9 >>
3.38 where 3.38 is the critical value obtained from Table 9
using n= 11 and v = 24 degrees of freedom. Since Table 9 skips
the desired value of n = 11, the larger of the neighboring
critical values, 3.38 for n =12 and 3.29 for n =10, is used). We
conclude that the procedures of Laboratory 12 should also be
investigated.
7.3.6 To verify that the remaining laboratories did indeed
obtain homogeneous results, we might repeat the analysis of
variance omitting Laboratories 10 and 12. The calculations
give the results shown in Table 12.
7.3.6.1 For this analysis, the variation between laboratories
is not significant at the 5 % level and we conclude that all the
laboratories except No. 10 and No. 12 exhibit the same
capability in testing procedure.
 E178 − 08
TABLE 9 Critical Values (One-Sided Test) for T ' When Standard Deviations v is Independent of Present SampleA
x n 2 x̄ x̄ 2 x 1
T' 5 , or
sv sv
n
v = d.f.
3 4 5 6 7 8 9 10 12
1 percentage point
10 2.78 3.10 3.32 3.48 3.62 3.73 3.82 3.90 4.04
11 2.72 3.02 3.24 3.39 3.52 3.63 3.72 3.79 3.93
12 2.67 2.96 3.17 3.32 3.45 3.55 3.64 3.71 3.84
13 2.63 2.92 3.12 3.27 3.38 3.48 3.57 3.64 3.76
14 2.60 2.88 3.07 3.22 3.33 3.43 3.51 3.58 3.70

15 2.57 2.84 3.03 3.17 3.29 3.38 3.46 3.53 3.65

16 2.54 2.81 3.00 3.14 3.25 3.34 3.42 3.49 3.60
17 2.52 2.79 2.97 3.11 3.22 3.31 3.38 3.45 3.56
18 2.50 2.77 2.95 3.08 3.19 3.28 3.35 3.42 3.53
19 2.49 2.75 2.93 3.06 3.16 3.25 3.33 3.39 3.50

20 2.47 2.73 2.91 3.04 3.14 3.23 3.30 3.37 3.47

24 2.42 2.68 2.84 2.97 3.07 3.16 3.23 3.29 3.38
30 2.38 2.62 2.79 2.91 3.01 3.08 3.15 3.21 3.30
40 2.34 2.57 2.73 2.85 2.94 3.02 3.08 3.13 3.22

60 2.29 2.52 2.68 2.79 2.88 2.95 3.01 3.06 3.15

ir
120 2.25 2.48 2.62 2.73 2.82 2.89 2.95 3.00 3.08
` 2.22 2.43 2.57 2.68 2.76 2.83 2.88 2.93 3.01
5 percentage points
10 2.01 2.27 2.46 2.60 2.72 2.81 2.89 2.96 3.08

a.
11 1.98 2.24 2.42 2.56 2.67 2.76 2.84 2.91 3.03
12 1.96 2.21 2.39 2.52 2.63 2.72 2.80 2.87 2.98
13 1.94 2.19 2.36 2.50 2.60 2.69 2.76 2.83 2.94
14 1.93 2.17 2.34 2.47 2.57 2.66 2.74 2.80 2.91

15
16
17
18
1.91
1.90
1.89
1.88
2.15
2.14
2.13
2.11
2.32
2.31
2.29
2.28
2.45
2.43
2.42
2.40
di 2.55
2.53
2.52
2.50
2.64
2.62
2.60
2.58
2.71
2.69
2.67
2.65
2.77
2.75
2.73
2.71
2.88
2.86
2.84
2.82
19 1.87 2.11 2.27 2.39 2.49 2.57 2.64 2.70 2.80
e
20 1.87 2.10 2.26 2.38 2.47 2.56 2.63 2.68 2.78
24 1.84 2.07 2.23 2.34 2.44 2.52 2.58 2.64 2.74
30 1.82 2.04 2.20 2.31 2.40 2.48 2.54 2.60 2.69
gP

40 1.80 2.02 2.17 2.28 2.37 2.44 2.50 2.56 2.65

60 1.78 1.99 2.14 2.25 2.33 2.41 2.47 2.52 2.61

120 1.76 1.96 2.11 2.22 2.30 2.37 2.43 2.48 2.57
` 1.74 1.94 2.08 2.18 2.27 2.33 2.39 2.44 2.52
A
The percentage points are reproduced from Ref (12).
En

7.3.6.2 In conclusion, there should be a systematic investi-
gation of test methods for Laboratories No. 10 and No. 12 to
determine why their test procedures are apparently different
from the other ten laboratories. (In this type of problem, the
tables of Greenhouse, Halperin, and Cornfield (13) could also
be used for testing outlying laboratory averages.)
7.3.7 Cautionary Remarks—In the use of the tests for
outliers as given above, our interest was to direct the statistical
tests of significance toward picking out those laboratories
which have different levels of measurement than the others.
Thus, we have assumed that there should not exist any
component of variance among the laboratory true means of
measurement. On the other hand, it is well known that in
practically all interlaboratory tests one does indeed find a
nonzero component of variance among the laboratory levels.
Often the variance among the laboratory means may be several
times that within individual laboratories. Thus, if we knew the
size of the actual component of variance among laboratories
we must live with—or guard against—then the observed F
ratio could be multiplied by the within variance of a sample
mean and divided by this quantity plus the among laboratory
variance, in order to adjust the F test to detect the undesirable
deviations of those laboratories which departed in average
level from measurements of the common or acceptable level of
the closely agreeing laboratories. Also, a somewhat similar
adjustment, if desired, could be applied to the tests for isolated
outliers. In our particular example, however, we desired to
detect those particular laboratories which departed in average
level from that of the closely agreeing laboratories. In fact, this
should be the aim of many interlaboratory testing programs, if
we are to seek high precision and accuracy of measurement.
8. Recommended Criteria for Known Standard Deviation
8.1 Frequently the population standard deviation σ may be
known accurately. In such cases, Table 13 may be used for
single outliers and we illustrate with the following example:

15
 E178 − 08
TABLE 10 Standardization of Sodium Hydroxide Solutions as TABLE 13 Critical Values (One-Sided Test) of T’1` andT' n` When
Determined by Plant Laboratories the Population Standard Deviation σ is KnownA
Standard used: Potassium Acid Phthalate (P.A.P.) Number of 5 Percent 1 Percent 0.5 Percent
Deviation Observations, Significance Significance Significance
(P.A.P. − n Level Level Level
Labora- of Average
0.096000 Sums Averages 2 1.39 1.82 1.99
tory from Grand
×10 3) 3 1.74 2.22 2.40
Average
1 1.893 4 1.94 2.43 2.62
1.972 5 2.08 2.57 2.76
1.876 5.741 1.914 + 0.043 6 2.18 2.68 2.87
2 2.046 7 2.27 2.76 2.95
1.851 8 2.33 2.83 3.02
1.949 5.846 1.949 + 0.078 9 2.39 2.88 3.07
3 1.874 10 2.44 2.93 3.12
1.792 11 2.48 2.97 3.16
1.829 5.495 1.832 −0.039 12 2.52 3.01 3.20
4 1.861 13 2.56 3.04 3.23
1.998 14 2.59 3.07 3.26
1.983 5.842 1.947 + 0.076 15 2.62 3.10 3.29
5 1.922 16 2.64 3.12 3.31
1.881 17 2.67 3.15 3.33
1.850 5.653 1.884 + 0.013 18 2.69 3.17 3.36
6 2.082 19 2.71 3.19 3.38
1.958 20 2.73 3.21 3.39
2.029 6.069 2.023 + 0.152 21 2.75 3.22 3.41
22 2.77 3.24 3.42

ir
7 1.992
1.980 23 2.78 3.26 3.44
2.066 6.038 2.013 + 0.142 24 2.80 3.27 3.45
8 2.050 25 2.81 3.28 3.46

a.
2.181 x1 # x2 # x3 # ... # xn
1.903 6.134 2.045 + 0.174 T'1 = (x̄ − x1)/σ; T'n = (xn − x̄ )/σ
9 1.831 A
This table is taken from Ref (13).
1.883
1.855 5.569 1.856 −0.015 di
10 0.735
0.722
0.777 2.234 0.745 −1.126 path. Since the stars were also photographed at the same times
11 2.064 as the Satellite, all the pictures show star-trails and are thus
1.794 called “star-plates.”
1.891 5.749 1.916 + 0.045
e
12 2.475 8.2.1 The x- and y-coordinate of each point on the Echo path
2.403 are read from a photograph, using a stereo-comparator. To
2.102 6.980 2.327 + 0.456
eliminate bias of the reader, the photograph is placed in one
gP

Grand
sum 67.350 position and the coordinates are read; then the photograph is
Grand rotated 180 deg and the coordinates reread. The average of the
average 1.871
two readings is taken as the final reading. Before any further
calculations are made, the readings must be “screened” for
TABLE 11 Analysis of Variance
gross reading or tabulation errors. This is done by examining
the difference in the readings taken at the two positions of the
En

Degrees
of
Sum of Mean photograph.
Source of Variation Squares Square F-ratio
Freedom
(SS) (MS)
8.2.2 Table 14 records a sample of six readings made at the
(d.f.) two positions and the differences in these readings. On the third
Between laboratories 11 4.70180 0.4274 F = v 48.61
Within laboratories 24 0.21103 0.008793 (highly significant) reading, the differences are rather large. Has the operator made
Total 35 4.91283 an error in placing the cross hair on the point?
8.2.3 For this example, an independent estimate of σ is
available since extensive tests on the stereo-comparator have
TABLE 12 Analysis of Variance shown that the standard deviation in reader’s error is about 4
(Omitting Labs 10 and 12)
µm. The determination of this standard error was based on such
Source of Variation d.f. SS MS F-ratio
Between laboratories 9 0.13889 0.015430 F = v 2.36
Within laboratories 20 0.13107 0.00655 F0.05(9, 20) = v 2.40 TABLE 14 Measurements, µm
F0.01(9, 20) = v 3.45
Total 29 0.26996 x -Coordinate y -Coordinate
Position Position
Position Position
1 + 180 ∆x 1 + 180 ∆y
1 1
deg deg
−53011 −53004 −7 70263 70258 +5
−38112 −38103 −9 −39739 −39729 −6
8.2 Example 7 (σ known)—Passage of the Echo I (Balloon) −2804 −2828 + 24 81162 81140 + 22
Satellite was recorded on star-plates when it was visible. 18473 18467 +6 41477 41485 −8
25507 25497 + 10 1082 1076 +6
Photographs were made by means of a camera with shutter 87736 87739 −3 −7442 −7434 −8
automatically timed to obtain a series of points for the Echo

16
 E178 − 08
a large sample that we can assume σ = 4 µm. The standard
deviation of the difference in two readings is therefore
=4 2 14 2 5 =32 or 5.7 µm (31)
8.2.4 For the six readings above, the mean difference in the
x-coordinates is ¯∆ x = 3.5 and the mean difference in the
y-coordinates is ¯∆ y = 1.8. For the questionable third reading,
we have
T' x 5 ~ 24 2 3.5! /5.7 5 3.60 (32)
T' y 5 ~ 22 2 1.8! /5.7 5 3.54 (33)
From Table 13 we see that for n = 6, values of T'n∞ as large
as the calculated values would occur by chance less than 1 %
of the time so that a significant reading error seems to have
been made on the third point.
8.3 A great number of points are read and automatically
tabulated on star-plates. Here we have chosen a very small
sample of these points. In actual practice, the tabulations would
probably be scanned quickly for very large errors such as
tabulator errors; then some rule-of-thumb such as 63 standard
deviations of reader’s error might be used to scan for outliers
due to operator error (Note 6). In other words, the data are
probably too extensive to allow repeated use of precise tests
like those described above (especially for varying sample size),
but this example does illustrate the case where σ is assumed
known. If gross disagreement is found in the two readings of a
coordinate, then the reading could be omitted or reread before
further computations are made.
NOTE 6—Note that the values of Table 13 vary between about 1.4σ and
e
3.5σ.
9. Additional Comments
gP
9.1 In the above, we have covered only that part of
screening samples to detect outliers statistically. However, a
large area remains after the decision has been reached that
outliers are present in data. Once some of the sample obser-
vations are branded as “outliers,” then a thorough investigation
En
should be initiated to determine the cause. In particular, one
should look for gross errors, personal errors, errors of
measurement, errors in calibration, etc. If reasons are found for
aberrant observations, then one should act accordingly and
perhaps scrutinize also the other observations. Finally, if one
reaches the point that some observations are to be discarded or
treated in a special manner based solely on statistical judgment,
then it must be decided what action should be taken in the
further analysis of the data. We do not propose to cover this
problem here, since in many cases it will depend greatly on the
particular case in hand. However, we do remark that there
could be the outright rejection of aberrant observations once
and for all on physical grounds (and preferably not on
statistical grounds generally) and only the remaining observa-
tions would be used in further analyses or in estimation
problems. On the other hand, some may want to replace
aberrant values with newly taken observations and others may
want to “Winsorize” the outliers, that is, replace them with the
next closest values in the sample. Also with outliers in a
sample, some may wish to use the median instead of the mean,
and so on. Finally, we remark that perhaps a fair or appropriate
practice might be that of using truncated-sample theory (11) for
cases of samples where we have “censored” or rejected some
of the observations. We cannot go further into these problems
here. For additional reading on outliers, see Refs (12,14,15,16,
17,18,19).
9.2 A sample test criterion for non-normality, and hence
ir
possibly for outliers, not covered above is the Wilk-Shapiro W
statistic for a sample of size n given by
F(@ n/2 #
G 2
a.
a n2i11 ~ x 2 x i!
n2i11
i51
W5 n (34)
di ( ~x
i51
i 2 x̄ ! 2
where:
x1 ≤ x2 ≤ x3 ≤ ... ≤ xn,
n
x̄5 ( x i /n,
i51
[n/2] is the greatest integer in n/2, and the coefficients an−i+1
are the order statistics for n = 2(1)50 given in Ref (20).
The Wilk-Shapiro W statistic has been found to be quite
sensitive to departures from normality and generally may
compare most favorably with the =b 1 and b2 tests discussed
above. In addition, therefore, the W statistic may also be used
as a test for outliers, or otherwise as a general test for
heterogeneity of sample values. Our significance tests given
above have been selected and recommended since they spe-
cifically point out particular suspected outliers in the sample.
We therefore are inclined to favor the above tests for specific
outliers in samples for the case where they will be used
routinely, for example, by engineers.
10. Keywords
10.1 dixon test; gross deviation; Grubbs test; outlier
17
 E178 − 08
REFERENCES

(1) Grubbs, F. E., and Beck, G.,“ Extension of Sample Sizes and
Percentage Points for Significance Tests of Outlying Observations,’’
Technometrics, TCMTA, Vol 14, No. 4, November 1972, pp. 847–854.
(2) Dixon, W. J., “Processing Data for Outliers,’’ Biometrics, BIOMA
March 1953, Vol 9, No. 1, pp. 74–89.
(3) Ferguson, T. S., On the Rejection of Outliers, Fourth Berkeley
Symposium on Mathematical Statistics and Probability, edited by
Jerzy Neyman, University of California Press, Berkeley and Los
Angeles, Calif., 1961.
(4) Ferguson, T. S., “Rules for Rejection of Outliers,’’ Revue Inst. Int. de
Stat., RINSA Vol 29, Issue 3, 1961, pp. 29–43.
(5) Grubbs, F. E., “Sample Criteria for Testing Outlying Observations,’’
Annals of Mathematical Statistics, AASTA Vol 21, March 1950, pp.
27–58.
(6) David, H. A., Hartley, H. O., and Pearson, E. S., “The Distribution of
the Ratio in a Single Sample of Range to Standard Deviation,
Biometrika, BIOKA Vol 41, 1954, pp. 482–493.
(7) Tietjen, G. L., and Moore, R. H., “Some Grubbs-Type Statistics for
the Detection of Several Outliers,’’ Technometrics, TCMTA, Vol 14,
No. 3, August 1972, pp. 583–597.
(11) Kudo, A., “On the Testing of Outlying Observations,” Sankhyā, The
Indian Journal of Statistics, SNKYA Vol 17, Part 1, June 1956, pp.
67–76.
(12) David, H. A., “Revised Upper Percentage Points of the Extreme
Studentized Deviate from the Sample Mean,’’ Biometrika, BIOKA
Vol 43, 1956, pp. 449–451.
(13) Greenhouse, S. W., Halperin, M., and Cornfield, J.,“ Tables of
Percentage Points for the Studentized Maximum Absolute Deviation
in Normal Samples,’’ Journal of the American Statistical
Association, JSTNA Vol 50, No. 269, 1955, pp. 185–195.
(14) Anscombe, F. J., “Rejection of Outliers,’’ Technometrics, TCMTA
Vol 2, No. 2, 1960, pp. 123–147.
(15) Chew, Victor,“ Tests for the Rejection of Outlying Observations,”
RCA Systems Analysis Technical Memorandum No. 64-7, 31 Dec.
1964, Patrick Air Force Base, Fla.
(16) Kruskal, W. H., “Some Remarks on Wild Observations,”
Technometrics, TCMTA Vol 2, No. 1, 1960, pp. 1–3.
(17) Proschan, F., “Testing Suspected Observations,” Industrial Quality
Control, IQCOA Vol XIII, No. 7, January 1957, pp. 14–19.
(18) Sarhan, A. E., and Greenberg, B. G., Editors, Contributions to Order

ir
(8) Chauvenet, W., A. Manual of Spherical and Practical Astronomy,Vol Statistics, John Wiley and Sons, Inc., New York 1962.
2, Fifth Edition. (19) Thompson, W. R. “On a Criterion for the Rejection of Observa-tions
(9) David, H. A., and Quesenberry, C. P., “Some Tests for Outliers,’’ and the Distribution of the Ratio of the Deviation to the Sample

a.
Technical Report No. 47, OOR(ARO) Project No. 1166, Virginia Standard Deviation,’’ The Annals of Mathematical Statistics,
Polytechnic Inst., Blacksburg, Va. AASTA Vol 6, 1935, pp. 214–219.
(10) Grubbs, F. E., “Procedures for Detecting Outlying Observations in (20) Shapiro, S. S., and Wilk, M. B., “An Analysis of Variance Test for
Samples,’’ Technometrics, TCMTA, Vol 11, No. 4, February 1969, Non-Normality (Complete Samples),” Biometrika, BIOKA, Vol 52,
pp. 1–21. 1965, pp. 591–611.
di
ASTM International takes no position respecting the validity of any patent rights asserted in connection with any item mentioned
in this standard. Users of this standard are expressly advised that determination of the validity of any such patent rights, and the risk
of infringement of such rights, are entirely their own responsibility.
e
This standard is subject to revision at any time by the responsible technical committee and must be reviewed every five years and
if not revised, either reapproved or withdrawn. Your comments are invited either for revision of this standard or for additional standards
gP

and should be addressed to ASTM International Headquarters. Your comments will receive careful consideration at a meeting of the
responsible technical committee, which you may attend. If you feel that your comments have not received a fair hearing you should
make your views known to the ASTM Committee on Standards, at the address shown below.

This standard is copyrighted by ASTM International, 100 Barr Harbor Drive, PO Box C700, West Conshohocken, PA 19428-2959,
United States. Individual reprints (single or multiple copies) of this standard may be obtained by contacting ASTM at the above
address or at 610-832-9585 (phone), 610-832-9555 (fax), or service@astm.org (e-mail); or through the ASTM website
(www.astm.org). Permission rights to photocopy the standard may also be secured from the Copyright Clearance Center, 222
En

Rosewood Drive, Danvers, MA 01923, Tel: (978) 646-2600; http://www.copyright.com/

18